Large Banking Systems with Default and Recovery: A Mean Field Game Model
ROMUALD ELIE ∗ TOMOYUKI ICHIBA † MATHIEU LAURIERE ‡
Abstract We consider a mean-field model for large banking systems, which takes into account default and recovery of the institutions. Building on models used for groups of interacting neurons, we first study a McKean-Vlasov dynamics and its evolutionary Fokker-Planck equation in which the mean-field interactions occur through a mean-reverting term and through a hitting time corresponding to a default level. The latter feature reflects the impact of a financial institution’s default on the global distribution of reserves in the banking system. The systemic risk problem of financial institutions is understood as a blow-up phenomenon of the Fokker-Planck equation. Then, we incorporate in the model an optimization component by letting the institutions control part of their dynamics in order to minimize their expected risk. Phrasing this optimization problem as a mean-field game, we provide an explicit solution in a special case and, in the general case, we report numerical experiments based on a finite difference scheme.
1 Introduction
Financial institutions form a highly connected network through monetary flow and complex de- pendencies. Each institution is trying to maximize its expected return objective over time, while the aggregation of all investment strategies generates feedback loops and results in some overall patterns of the financial market. The institutions are all competing against each other as players in a financial interacting game. When the number of players becomes large as we can observe in the current fully connected worldwide banking system, individual interactions become intractable while the global patterns become more apparent. Our goal in this paper is to capture the origins of these patterns in a simple mathematical and numerical setup, described by a mean-field game banking system, taking into account births and defaults of financial institutions. The set up of the birth and default dynamics considered here is inspired by the neuron firing arXiv:2001.10206v1 [math.OC] 28 Jan 2020 models [3, 9, 8, 17] as well as mathematical physics models [25, 26]; see also [12, 13, 19, 23]. These models aim at describing how the electric potential of a network of neurons evolves over time. A salient feature occurs whenever the potential of one single neuron reaches a given threshold level, leading to a so-called firing event where all the energy accumulated is transmitted to neighboring neurons in the network. In these models, it has been observed that a few parameters characterize some sort of phase transition period, according to which steady states or blow-up phenomena may occur or not. A blow-up typically corresponds to the situation in which the firing of one neuron
∗ Université Paris-Est Marne-la-Vallée, France (E-mail: [email protected]). † Department of Statistics and Applied Probability, South Hall, University of California, Santa Barbara, CA 93106, USA (E-mail: [email protected]). Research supported in part by the National Science Foundation under grants NSF-DMS-13-13373 and DMS-1615229 ‡ Department of Operations Research and Financial Engineering, Sherrerd Hall, Princeton University, Princeton, NJ 08540, USA (E-mail: [email protected]).
1 immediately triggers the firing of some other neurons and so on, resulting in a cascade of firings. In the macroscopic limit, this translates into a jump in the instantaneous rate of firings. Handling in a mathematically rigorous way such phenomenon is an active and very challenging research area, in which recent progress has been made for instance in [10] in connection with the aforementioned line of work. Drawing an analogy between neurons firing and banks defaulting, the blow-up phenomenon appears as a natural tool for describing systemic patterns of a financial crisis: complicated interac- tions between poorly regulated banks driven by selfish objectives can sometimes lead to cascade of defaults. This line of modeling has already been pointed out e.g. in [7, 26] to describe the evolution of a financial system. Building on this type of model, our main contribution in this paper is to incorporate a proper optimization component for each bank, by letting each bank influence the dynamics of its state so as to minimize a chosen idiosyncratic risk criterion, in the spirit of [6].
More specifically, we focus here on the mean-field limit of the following toy model: The banking system consists of N banks, where each bank i owns liquid assets as monetary reserve evolving i continuously in time. Let denote by (Xt )t≥0 the non-negative level of cash reserve (in liquid assets) of bank i , for i = 1,...,N and introduce (Xt)t≥0 the average level of cash reserves over the full 1 N banking system, i.e., Xt := (Xt + ··· + Xt ) /N at time t ≥ 0 . We assume for simplicity that the diffusion dynamics of each cash reserve Xi is of mean field type, i.e. it depends on the global banking system through the empirical distribution of the cash reserve levels, and more specifically through their empirical average X . More precisely, the diffusive i 2 i behavior of X at time t is locally specified by a drift function b : R+ → R of its value Xt together i with the running average Xt of the system. The sample path t → Xt is assumed to be right continuous with left limits. i0 The bank defaults are modeled in the following way: if the amount Xt of liquid assets of bank i0 reaches the threshold level 0 at time t0 , this bank i0 is defaulted from the system. A systemic effect induces a financial shock to each institution of the system, so that the cash reserve Xj of every bank j(=6 i0) suffers at time t0− a downward jump of size Xt0−/N . Instantaneously and for ease of modeling, a new institution is also created in the market with cash reserve at the average level Xt0−, so that the number of banks in the financial system remains constant. For notational simplicity, this new bank keeps the same number , i.e., i0 i0 . With initial i0 Xt0− = 0 ,Xt0+ := Xt0 1 N N configuration x0 := (X0 ,...,X0 ) ∈ (0, ∞) , the resulting dynamics may be depicted by the following system of stochastic differential equations
Z t Z t i i i i i 1 X j X = X + b(X , Xs)ds + σW + Xs− dM − dM ; t ≥ 0 , t 0 s t s N s 0 0 j6=i ∞ (1) i X i n i i Xs− X j j o Mt := 1{τ i ≤t} , τk := inf s > τk−1 : Xs− − (Ms − Ms−) ≤ 0 ; k ∈ N , k N k=1 j6=i for i = 1,...,N , where W := (W 1,...,W N ) , t ≥ 0 is a standard N-dimensional Brownian mo- 1 N i tion, X is the average of X := (X ,...,X ) , Mt is the cumulative number of defaults of bank i i i by time t ≥ 0 , and τk is its k -th default time with τ0 = 0 . Although, in our toy model, the drift coefficient b depends on both the state X of a bank together with the average X whereas the dif- fusion coefficient is a positive constant σ, one can in general consider more complicated dependence on the coefficients, such as time-dependent drift and diffusion coefficients. The system (1) of N banks includes mean-field interactions through the drift function b(·) as well as via the cumulative number of defaults M. The mathematical analysis of such system raises some delicate and technical
2 issues, due to mean-field interactions together with multiple defaults.
We then turn our attention to the more realistic situation in which banks are not passively following the dynamics (1) but are able to partly control their drifts and seek to minimize the sum of expected running costs (possibly together with a terminal default cost), occurring at default time. Compared with the model (1), the drift of bank i incorporates a linear dependence on the control i ξt used by this bank. The running cost rate f at time t ≥ 0 of bank i depends on the control i i PN ξ , the monetary reserve X as well as the empirical distribution mN,t(·) := δ i (·) /N of t t i=1 Xt the system. This allows in particular to penalize strong deviations from the average of the banking system. Here, we borrow the running cost functional from a model introduced in [6] for systemic risk: The running cost functional takes the form of a quadratic function. Moreover, the running cost is discounted at rate r . Thus, over the time interval [0,T ] , each bank i seeks to minimize
"Z τ i∧T # −rs i i E e f(Xs, mN,s, ξs)ds , 0 where X is controlled by ξ via its drift. In this context, we look for a Nash equilibrium, in the sense that when bank i performs the optimization, the controls used by the other banks j =6 i are fixed. This modeling of systemic risk hence identifies to an N -player stochastic game, as the other players policy and default rate modify the banking system dynamics, under which each bank optimizes its policy and resulting monetary reserve. As the size of the system becomes large, i.e., N → ∞ , the difficulty and complexity of the mathematical analysis increase rapidly. In order to obtain a tractable approximation of a large banking system, we rely on the mean field game (MFG) paradigm. Mean field games were introduced by Lasry and Lions [24, 22, 20, 21] and Caines, Huang and Malhamé [14, 15, 16]. The interested reader is referred to the recently published books [2, 4, 5] and the references therein.
The rest of the paper is organized as follows. In section 2, we first show that the N-agent system (1) is well defined using a notion of physical solution borrowed from [9]. We then informally derive the limiting mean-field system as N → ∞ , as well as the corresponding nonlinear Fokker- Plank equation. In the limiting system, we show in Theorem 4 that blow-up may occur when the initial distribution is concentrated near the origin. This result shades in particular a new light on the blow-up phenomenon described in [3]. Importantly, we observe that the Fokker-Plank system has an explicit stationary solution, as derived in Theorem 5. We conclude section 2 with numerical results for the dynamics of the system. In section 3, we incorporate the optimization component. We formulate the mean-field game in section 3.1. Its solution is characterized by a Hamilton-Jacobi- Bellman (HJB) equation coupled with a Fokker Plank (FP) equation with well suited boundary conditions at the boundary of the domain. In section 3.3, we derive an explicit solution for a stationary mean field game from the PDE system, where Theorem 5 is used to determine the probability distribution. This stationary solution serves as a baseline for our numerical study. In section 3.4, we provide a discrete numerical scheme for the HJB-FP system, building on [1]. Finally, numerical results in a non-stationary regime are presented in section 3.5.
3 2 Fokker Planck equation for the particle system
2.1 Construction of Physical Solution 2.1.1 Modeling bank interactions and defaults Throughout this section, we assume for simplicity that σ = 1, and in this subsection, let us assume 2 that b(·, ·) in (1) is (globally) Lipschitz continuous on R+ , i.e., there exists a constant κ > 0 such that |b(x1, m1) − b(x2, m2)| ≤ κ |x1 − x2| + |m1 − m2| (2) for all x1, x2, m1, m2 ∈ R+ , and impose the following condition on the drift function b(·, ·) :
N X b(xi, x) ≡ 0 (3) i=1
1 N N 1 N for every x := (x , . . . , x ) ∈ R+ and x := (x + ··· + x ) /N . This condition holds for instance if b(xi, x) = xi − x is a linear mean-reverting drift, and naturally translates a global stability of the monetary level for the whole financial system.
1 N Given a standard Brownian motion W· , we shall consider a system (X· := (X· ,...,X· ),M· := 1 N (M· ,...,M· )) described by (1) together with conditions (2)-(3) on a filtered probability space (Ω, F, F, P) , with filtration F := (Ft, t ≥ 0) . In particular, we are concerned with identifying cases where dynamics (1) might induce multiple simultaneous defaults with positive probability, i.e.,